Solar Mini 4 vs Solar Pro 4: the Same Context Window at Three Times the Price

Upstage's two current Solar models share a spec sheet on every dimension a buyer normally checks first. Solar Mini 4, released 22 September 2026, and Solar Pro 4, released 6 August 2026, both accept 512,000 tokens of context, both emit up to 128,000 output tokens, both cover English, Korean and Japanese, and both stop their training data in February 2026. The difference is that Solar Pro 4 lists at three times Solar Mini 4's price — $0.30 per million input tokens against $0.10 — and that Solar Pro 4 has a published independent benchmark score while Solar Mini 4, one day old, has none at all. Which means the question this comparison answers is not "which is better." It is whether the thing you are paying three times for is the thing your workload is short of.

Two discount clocks running at once

Before the capability argument, the price argument — because right now the two models are on different promotional timelines and the arithmetic is stranger than the list prices suggest.

Solar Mini 4 launched on 22 September with 50% off through 22 October 2026, which puts its current rates at $0.05 input, $0.005 cached input and $0.20 output per million tokens. Upstage's own pricing page carries the structured window as 22 September 05:00 UTC through 23 October 00:00 UTC.

Solar Pro 4's promotion moved on 11 September, when Upstage cut the discount from 90% to 70% — a price increase, in plain terms, though the model stayed well below list. That promotion runs until 9 October 2026, with standard pricing from 10 October. During it, Solar Pro 4 sits at $0.09 input, $0.018 cached and $0.36 output.

So for the roughly two weeks between now and 9 October, the gap between the two models is 1.8×, not 3×: $0.09 against $0.05 on input, $0.36 against $0.20 on output. After 9 October, Solar Pro 4 reverts to $0.30 and $1.20 while Solar Mini 4 stays discounted until the 22nd. After the 22nd, both are at list and the gap is the full 3×.

Three conclusions fall out of that. If you are benchmarking the two against each other, do it before 9 October, when the cheaper model's advantage is smallest and you are measuring capability rather than price. If you are budgeting a production rollout that starts in November, use the list prices, because neither promotion will still be running. And if you are planning a migration between the two, know that the crossover point is a date, not a workload.

The one place the two models are genuinely far apart

Independent evaluation is where Solar Pro 4 earns its tier and Solar Mini 4 cannot yet answer the question.

Artificial Analysis scored Solar Pro 4 at 42 on its Intelligence Index, a 27-point jump over Solar Pro 3's 14, and the gains were concentrated exactly where a flagship should win. Terminal-Bench v2.1 went from 12% to 57%. AA-LCR, the long-context reasoning evaluation, went from 31% to 71%. τ³-Banking went from 9% to 23%. On GDPval-AA v2, which measures real-world agentic tasks against a human Elo baseline of 1000, Solar Pro 3 scored 498 and Solar Pro 4 scored 1277. Those are third-party figures, published by Artificial Analysis, not vendor numbers.

The same write-up is honest about the cost of that improvement. Solar Pro 4 takes 8.6 minutes to complete an average Intelligence Index task against Solar Pro 3's 6.0, despite using about 17% fewer output tokens — 43,000 per task against 52,000. It is more token-efficient and slower in wall-clock terms. And its AA-Omniscience score improved from −53 to −1 mostly by abstaining: it attempts only 41% of questions against Pro 3's 92%, and while its hallucination rate fell from 88% to 24%, its accuracy barely moved, from around 17% to 19%. A model that improves its honesty score by declining to answer is a real improvement for agent work — a confident wrong answer costs more than a refusal — but it is not the same thing as knowing more.

Solar Mini 4 has no equivalent. No Intelligence Index, no Agent Arena rank, no third-party agentic evaluation, and no benchmark table from Upstage either. The only figure in circulation is a community wrapper's self-reported test400 run judged by another model, in which Solar Mini 4 scored 98.4% field accuracy with a 1.21-second average call time. That is one developer's harness against their own implementation, useful as a smoke signal and useless as a procurement input.

What 3B active parameters buys, and does not

The parameter disclosure is the clearest statement of intent in this pairing. Solar Mini 4 is 35B total with 3B active per token. Solar Pro 4's parameter count is undisclosed.

A 3B active ratio is a serving-economics decision, not a capability claim. It means the model stores 35B of weights and does the arithmetic of 3B for every token, which is why Upstage can price a 512K-context model at $0.10 per million input tokens. Solar Pro 4 at $0.30 tells you it is doing more work per token; Upstage not publishing the number tells you it does not want the comparison made arithmetically.

What actually separates them, based on the published evidence, is the difficulty of the task the model is asked to finish rather than the size of the document it is handed. Solar Mini 4 was announced with 512K context and up to 128K output — the same figures as the flagship — so the long-document case is not where the tier boundary sits. The boundary is at multi-step agentic work: the Terminal-Bench and τ³-Banking and GDPval numbers are what Solar Pro 4's price is buying, and they measure tasks that require planning, tool calls and recovery from mistakes over many turns.

There is a caveat worth stating, because it is the kind of thing that bites people mid-project. Upstage's model history lists Solar Pro 4 at 512K context with 128K maximum output. Artificial Analysis's write-up on the same model lists a 384K context window and a 256K maximum output — the same ceiling and floor, swapped. Both sources are reputable and they disagree. If you are sizing a prompt near the limit, verify against the API response rather than either document, because the two published figures are not reconcilable.

Where each one belongs in a pipeline

The useful way to hold these two models is as a tier pair rather than as alternatives, which is roughly how Upstage has positioned them. Solar Mini 4 is the default: extraction, classification, structured output over long documents, the high-volume loop where a million cheap calls is the workload. Solar Pro 4 is the escalation: the agent run that has to plan, use tools, recover from a failed step and still finish.

The economics of that split are unforgiving in favour of the cheap tier whenever the cheap tier is right. Solar Mini 4 at $0.05 input and $0.20 output during its promotion is five to six times cheaper than Solar Pro 4's promotional rate and three times cheaper at list. If 90% of your volume is extraction and 10% is agentic work, routing the 90% down a tier saves more than any prompt optimisation will. The failure mode is misrouting — sending agentic work to the small model and paying for it twice, once in the wasted call and once in the retry on the expensive one.

That misrouting risk is the reason this is a measurement problem before it is a pricing problem, and it is where a routing layer earns its place. To be straight about the product boundary: OrcaRouter does not route any Upstage model, so neither Solar Mini 4 nor Solar Pro 4 is available through us — both come from Upstage's own API and several third-party platforms. What the platform is for is the pattern this pairing creates: a cheap tier and an expensive tier, a rule for which traffic goes where, and automatic failover that promotes a call to the larger model when the smaller one errors or returns something your parser rejects. Across the 190-odd models we do route, that tiering is configured once rather than wired into every caller. If you are running both Solar tiers, the equivalent logic lives in Upstage's own console, and the pieces you would want to read carefully are the ones that decide when to escalate.

What to do with them

If you have to choose one today, choose Solar Pro 4 only if your work is agentic in the specific sense the benchmarks measure — multi-step, tool-using, self-correcting. That is what the 3× is paying for, and it is the only axis where the published evidence clearly separates them. For everything else, Solar Mini 4 offers the same context window, the same output ceiling, the same three languages and a lower price, and the honest objection to it is not that it is weak but that nobody outside Upstage has measured it yet.

That last point is the whole decision, and it has a deadline attached. Solar Mini 4's promotion runs to 22 October and Solar Pro 4's to 9 October, so the cheap window for comparing them against each other is the next two weeks. Run both on your own traffic at the promotional rates, with your own acceptance criteria, before either clock runs out. Independent benchmarks for Solar Mini 4 will arrive eventually, and when they do they will tell you how it performs on someone else's task, which is a different and less useful answer than the one you can buy for a few dollars this month.